8/12/2019 Does Malware Detection Improve With Diverse
http://slidepdf.com/reader/full/does-malware-detection-improve-with-diverse 1/13
Cukier, M., Gashi, I., Sobesto, B. & Stankovic, V. (2013). oes Ma!"are etection I#$rove %ith
iverse ntiVirus 'roucts n *#$irica! Stu+. 'a$er $resente at the 32n Internationa!
Conerence on Co#$uter Saet+, -e!iabi!it+ an Securit+ (S*C/M'), 2 2 Se$te#ber 2013,
ou!ouse, rance.
Cit+ -esearch /n!ine
Original citation4 Cukier, M., Gashi, I., Sobesto, B. & Stankovic, V. (2013). oes Ma!"are etection
I#$rove %ith iverse ntiVirus 'roucts n *#$irica! Stu+. 'a$er $resente at the 32n
Internationa! Conerence on Co#$uter Saet+, -e!iabi!it+ an Securit+ (S*C/M'), 2 2
Se$te#ber 2013, ou!ouse, rance.
Permanent City Research Online URL4 htt$455o$enaccess.cit+.ac.uk523365
Copyright & reuse
Cit+ 7niversit+ 8onon has eve!o$e Cit+ -esearch /n!ine so that its users #a+ access the
research out$uts o Cit+ 7niversit+ 8onon9s sta. Co$+ri:ht ; an Mora! -i:hts or this $a$er are
retaine b+ the iniviua! author(s) an5 or other co$+ri:ht ho!ers. !! #ateria! in Cit+ -esearch
/n!ine is checke or e!i:ibi!it+ or co$+ri:ht beore bein: #ae avai!ab!e in the !ive archive. 7-8s
ro# Cit+ -esearch /n!ine #a+ be ree!+ istribute an !inke to ro# other "eb $a:es.
Versions of research
he version in Cit+ -esearch /n!ine #a+ ier ro# the ina! $ub!ishe version. 7sers are avise to
check the 'er#anent Cit+ -esearch /n!ine 7-8 above or the status o the $a$er.
Enquiries
I +ou have an+ en<uiries about an+ as$ect o Cit+ -esearch /n!ine, or i +ou "ish to #ake contact
"ith the author(s) o this $a$er, $!ease e#ai! the tea# at $ub!ications=cit+.ac.uk.
8/12/2019 Does Malware Detection Improve With Diverse
http://slidepdf.com/reader/full/does-malware-detection-improve-with-diverse 2/13
Does Malware Detection Improve With Diverse
AntiVirus Products? An Empirical Study
Michel Cukier 1, Ilir Gashi2, Betrand Sobesto1, Vladimir Stankovic2
1 University of Maryland, College Park, MD, USA
{mcukier, bsobesto}@umd.edu2Centre for Software Reliability, City University London, London, UK
{i.gashi, v.stankovic}@csr.city.ac.uk
Abstract. We present results of an empirical study to evaluate the detection ca-
pability of diverse AntiVirus products (AVs). We used malware samples col-
lected in a geographically distributed honeypot deployment in several different
countries and organizations. The malware was collected in August 2012: the re-
sults are relevant to recent and current threats observed in the internet. We sent
these malware to 42 AVs available from the VirusTotal service to evaluate the
benefits in detection from using more than one AV. We then compare these
findings with similar ones performed in the past to evaluate diversity with AVs.
In general we found that the new findings are consistent with previous ones, de-
spite some differences. This study provides additional evidence that detection
capabilities are improved by diversity with AVs.
Keywords. Empirical study; Intrusion tolerance; Malware; Measurement tech-niques; Security; Security assessment tools.
1 Introduction
All systems need to be sufficiently reliable and secure in delivering the service that
is required of them. Various ways in which this can be achieved in practice: from the
use of various validation and verification techniques; to the use of software
fault/intrusion tolerance techniques; or continuous maintenance and patching once the
product is released. Fault tolerance techniques range from simple “wrappers” of the
software components [1] to the use of diverse software products in a fault-tolerant
system [2]. Implementing fault tolerance with diversity was historically considered
prohibitively expensive, due to the need for multiple bespoke software versions.
However, the multitude of available off-the-shelf software for various applications
has made the use of software diversity an affordable option for fault tolerance against
either malicious or accidental faults.
Authors in [3] detailed an implementation of an AntiVirus (AV) platform that
makes use of diverse AVs for malware detection. A similar architecture that uses
diverse AV email scanners has been commercially available for several years [4].
8/12/2019 Does Malware Detection Improve With Diverse
http://slidepdf.com/reader/full/does-malware-detection-improve-with-diverse 3/13
Thus, architectural solutions for employing diverse AV detection engines are already
known and even commercially deployed. Results from empirical evaluation of the
effectiveness of diversity for malware detection are, however, much more scarce.
The following claim is made on the VirusTotal site [5]: “Currently, there is no so-
lution that offers 100% effectiveness in detecting viruses, malware and malicious
URLs”. Given these limitations of individual AVs, designers of security protect ion
systems are interested in at least getting estimates of the possible gains in terms of
added security that the use of diversity (e.g. diverse AVs) may bring for their systems.
Two of the authors of this paper have previously reported [6-8] results from a study
on the detection capabilities of different AVs and potential improvements in detection
that can be observed from using diverse AVs. In those studies we reported that some
AVs achieved high detection rates, but none detected all the malware samples. Wealso found many cases of regression in the detection capability of the AVs: cases
where an AV would regress from detecting the malware on a given date to not detect-
ing the same malware at a later date(s). We reported significant improvements in the
detection capability when using two or more diverse AVs. For example, even though
no single AV detected all the malware in these studies, almost 25% of all the diverse
1-out-of-2 pairs of AVs successfully detected all the malware.
The results presented in [6-8] are intriguing. However, they concern a specific
snapshot in the detection capabilities of AVs against malware threats prevalent in that
time period: 1599 malware samples collected from a distributed honeypot deployment
over a period of 178 days from February to August 2008. In the security field the
threat landscape changes rapidly and it is not clear to what extent these findings can
be generalized to currently spreading malware. It is also not clear whether the diversi-ty benefits reported in [6-8] are specific to that specific collection environment and
time period, or whether they are consistent with other environments and time periods.
Our work is motivated by the following claim from Fred Schneider in [9]: “ Exper-
imentation is the way to gain confidence in the accuracy of our approximations and
models. And just as experimentation in the natural sciences is supported by laborato-
ries, experimentation for a science of cybersecurity will require test beds where con-
trolled experiments can be run.” In this paper we present results of an empirical study
about possible benefits of diversity with currently spreading malware and compare
our findings with those reported in [6-8]. The main aim of our study is to verify the
extent to which the findings previously reported are relevant with more recent mal-
ware. Consistent with the statement in [9], through experimentation and empirical
analysis our goal is to gain confidence in the accuracy of the claims and help security
decision makers and researchers to make more informed, empirically-supported deci-
sions about the design, assessment and deployment of security solutions.
The results provide an interesting analysis of the detection capability of the respec-
tive signature-based components, though more work is needed for assessing full de-
tection capabilities of the AVs. Also, the purpose of our study is not to rank the indi-
vidual AVs, but to analyze the effectiveness of using diverse AVs.
8/12/2019 Does Malware Detection Improve With Diverse
http://slidepdf.com/reader/full/does-malware-detection-improve-with-diverse 4/13
The rest of the paper is organized as follows: Section 2 summarizes related work;
Section 3 describes the data collection infrastructure; Section 4 presents the results of
our study; Section 5 compares the results reported in this paper with the ones in [6-8];
Section 6 discusses the implications of our results on the decision making about secu-
rity protection systems, and presents conclusions and possible further work.
2 Related Work
Studies which perform analysis of the detection capabilities and rank various AVs
are very common. One such study, which provides analysis of “at risk time” for single
AVs, is given in [10]. Several sites1 report rankings and comparisons of AVs, though
readers should be careful about the definitions of “system under test” when compar-ing the results from different reports.
Empirical analyses of the benefits of diversity with diverse AVs are much less
common. Apart from our previous work [6-8] (we refer to the findings reported there
when we compare them with the new findings in Section 5) we know of only one
other published study [3] that has looked at the problem. An initial implementation of
the Cloud-AV architecture has been provided in [3], which utilizes multiple diverse
AVs. The Cloud-AV uses the client-server paradigm. Each machine in a network runs
a host service which monitors the host and forwards suspicious files to a centralized
network service. This service uses a set of diverse AVs to examine the file, and based
on the adopted security policy makes a decision regarding maliciousness of the file.
This decision is then forwarded to the host. To improve performance, the host service
adds the decision to its local repository. Hence, subsequent encounters of the samefile by the host will be decided locally. The implementation from [3] handles execut-
able files only. A study with a Cloud-AV deployment in a university network over a
six month period is given in [3]. For the files observed in the study, the network over-
head and the time needed for an AV to make a decision are relatively low. This is
because the processes running on the local host, during the observation period, could
make a decision in more than 99% of the cases. The authors acknowledge that the
performance penalties could be much higher if file types other than just executables
are examined, or if the number of new files observed on the host is high (since the
host will need to forward the files for examination to the network service more often).
Apart from Cloud-AV, which is an academic prototype, commercial solutions that
use diverse AV engines for file and e-mail scanning are also available2.
3 Experimental Infrastructure
The malware have been collected on a distributed honeypot architecture using
Dionaea - a low-interaction honeypot used to emulate common vulnerabilities, cap-
1 av-comparatives.org , av-test.org/, virusbtn.com/index2 gfi.com/maildefense/ , pcworld.com/article/165600/G_data_internet_security.html
8/12/2019 Does Malware Detection Improve With Diverse
http://slidepdf.com/reader/full/does-malware-detection-improve-with-diverse 5/13
ture malicious payloads attempting to exploit the vulnerabilities and collect the binary
files downloaded or uploaded. In summary, the main components of the experimental
infrastructure and the process of data collection are as follows (full details are availa-
ble in our technical report [11]):
Dionaea has been deployed on 1136 public IP addresses distributed in six different
locations in France, Germany, Morocco and the USA.
The subnets do not contain the same number of IP addresses and the configuration
differs between networks. Many IP addresses belong to University of Maryland
(which is where two of the authors of this paper are based). Note that neither all
networks apply the same security policies nor are protected in the same way.
We deployed the default configuration of Dionaea which exposes common Internet
services such as http, ftp, smtp, MS SQL, MySQL, as well as Microsoft Windowsand VOIP protocols. These services and protocols emulate known vulnerabilities
and can trap malware exploiting them. Due to the types of vulnerabilities and pro-
tocols emulated, Dionaea mainly collects Windows Portable Executable (PE) files.
For this study, one Dionaea instance was deployed on each separate subnet, run-
ning on a different Linux virtual machine. To facilitate the analysis, all the mal-
ware collected by Dionaea and the information relative to their submission were
merged and centralized on a single server.
Every day at midnight a Perl script downloads from the virtual machines running
Dionaea the binary files and the SQLite database containing the malware capture
information. This script then submits the whole malware repository to VirusTotal
[5], which is a web service providing online malware analysis based on several
AVs. For each binary file successfully submitted, VirusTotal returns a scan key.
The scan key is composed of the binary’s SHA1 hash and the timestamp of the
submission. To ensure a correct submission of each file and to later retrieve the
analysis results the script keeps track of the scan keys.
A second Perl script retrieves the reports for each malware using the scan keys
generated by VirusTotal. VirusTotal returns an array containing: the number of
AVs that have flagged the file as malicious, the total number of AVs used in the
analysis, and the AVs names, versions and the signatures name.
4 Results
4.1 Basic Statistics from Our Results
Using the dataset introduced in Section 3 we explore the potential benefits in mal-
ware detection from employing diverse AVs. We start with some descriptive statistics
of the obtained results.
Data collection lasted for 23 days: 8-30 August 2012. During this period we col-
lected 922 malware. These malware were sent to VirusTotal where they were exam-
ined by up to 42 AVs. We sent the malware on the first day of observation and con-
8/12/2019 Does Malware Detection Improve With Diverse
http://slidepdf.com/reader/full/does-malware-detection-improve-with-diverse 6/13
tinued to send them throughout the collection period. However the total number of
data points we have is not simply 23 * 922 * 42. It is smaller because:
not all malware were observed in the first day of collection – we continued to ob-
serve new malware throughout the collection period and we cannot send a newly
collected malware to older versions of AVs running on VirusTotal;
VirusTotal may not always return results for all AVs – we are not sure why this is.
VirusTotal is a black-box service and its internal configuration is not provided. We
presume that each AV is given a certain amount of time to respond; if it doesn’t,
VirusTotal will not return a result for that AV. Additionally a particular AV may
not be available at the time we submit the malware for inspection.
A unique “demand” for the purpose of our analysis is a { Malware j, Datek } pair which
associates a given malware j to a given date k in which it was sent to VirusTotal. We
treat each of the malware sent on a different date as a unique demand. If all 922 mal-
ware were sent to VirusTotal on each of the 23 days of data collection, then we would
have 922 * 23 = 21,206 demands. But as explained before, due to missing data, the
number of demands sent to any of the AVs is smaller than 21,206.
If we now associate a given AV i’s response to a given malware j on a given date k
then we can consider each of our data points in the experiment to be a unique triplet
{AVi, Malware j, Datek }. For each triplet we have defined a binary score: 0 in case of
successful detection, 1 in case of failure. Table 1 shows the aggregated counts of the
0s and 1s for the whole period of our data collection. We have considered as success
the generation of an alert by an AV regardless of the nature of the alert itself.
Table 1 - Counts of detetcions and failures for triplets {AVi, Malware j, Datek }
Value Count
0 – detection / no failure 766,853
1 – no detection / failure 65,410
4.2 Single AV Results
Table 2 contains the failure rates of all the 42 AVs. The ordering is by the failure
rate (second column) with the AV with the smallest failure rate appearing first.
The third column in Table 2 counts the number of “releases” of a given AV rec-
orded by VirusTotal. We presume these are the versions of either the rule set or the
release version of the detection engine itself. It seems that different products havedifferent conventions for this. Amongst the three equally best AVs in our study,
Ikarus reports only one version whereas AntiVir and ESET-NOD32 have 36 and 27
sub-release versions respectively.
The fourth and fifth columns of Table 2 report on an interesting phenomenon first
reported in our previous work [8] – some AVs regress on their detection capability.
That is, they detected the malware at first, and then failed to detect the malware at a
8/12/2019 Does Malware Detection Improve With Diverse
http://slidepdf.com/reader/full/does-malware-detection-improve-with-diverse 7/13
Table 2. Failure Rates for Each AV
AV Name Failure rate Number of “releases” of
the AV in VirusTotal
Count of Malware on
which AV regressed
Count of Regression
Instances
AntiVir 0.000049 1 0 0
ESET-NOD32 0.000049 36 0 0
Ikarus 0.000049 27 0 0
Kaspersky 0.000050 1 1 1
Sophos 0.000050 2 0 0
VIPRE 0.000098 29 0 0
McAfee 0.000099 1 0 0
Norman 0.000099 1 0 0
Emsisoft 0.000208 1 3 3
Symantec 0.001112 2 0 0
F-Secure 0.001204 1 0 0
Avast 0.001211 1 0 0
BitDefender 0.001232 1 0 0PCTools 0.001244 1 0 0
Jiangmin 0.001286 1 0 0
AVG 0.001342 1 0 0
GData 0.001627 1 8 8
TrendMicroHouseC. 0.001847 2 13 13
K7AntiVirus 0.001881 19 0 0
McAfee-GW-Ed. 0.002232 1 43 43
VirusBuster 0.002256 27 0 0
nProtect 0.002503 26 0 0
Microsoft 0.002515 3 0 0
TheHacker 0.002522 1 1 1
TrendMicro 0.002530 2 26 26
DrWeb 0.003517 2 0 0
ViRobot 0.003518 1 0 0
Panda 0.003530 1 4 17
TotalDefense 0.003708 23 0 0VBA32 0.006567 2 55 55
Comodo 0.009233 32 139 151
CAT-QuickHeal 0.010306 1 1 1
AhnLab-V3 0.010950 25 105 120
F-Prot 0.011789 1 1 1
Commtouch 0.012832 1 1 1
eSafe 0.165981 1 8 9
Rising 0.226335 18 10 10
SUPERAntiSpyware 0.356355 2 0 0
ClamAV 0.377830 1 0 0
Antiy-AVL 0.412142 1 1 1
Fortinet 0.670168 1 5 5
ByteHero 0.963825 1 33 59
later date(s), probably due to some updates in the respective AV’s rule definitions.
The fourth column contains the number of malware on which a given AV regressed,and the fifth column contains the number of instances of these regressions (since an
AV may have regressed more than once on a given malware: e.g. alternated between
detection and non-detection of a malware several times). We note that even a few
AVs, which are in the top ten in terms of the overall detection rates, did have cases of
regressions. Such a phenomenon can be due to various reasons. For instance, the ven-
dor might have deleted the corresponding detection signature as a consequence of the
identification of false positives associated to it, or they might be attempting to consol-
8/12/2019 Does Malware Detection Improve With Diverse
http://slidepdf.com/reader/full/does-malware-detection-improve-with-diverse 8/13
8/12/2019 Does Malware Detection Improve With Diverse
http://slidepdf.com/reader/full/does-malware-detection-improve-with-diverse 9/13
Of course many other voting configurations are possible, but we chose these two
above as they best represent the contrasting tradeoffs between detection rates and
false alarm rates that decision makers should take into consideration when deciding
on a diverse system configuration: a 1ooN configuration could be used to maximize
the detection rates of genuine malware; a majority voting one could be used to curtail
the false positive rate; the trigger level configurations allow a trade-off between these
two. Note that since we are dealing with confirmed malware samples we cannot really
present any data about false positives. However, the rooN results allow us to get ini-
tial estimates of how is the rate of correct detections of confirmed malware affected if
one uses majority voting or trigger-level setup to reduce the false positive rate.
Due to space limitations we only present a snapshot of the results we obtained. We
concentrated on showing a graphical representation of the results. The details and thetables from which these graphs are generated as well as other detailed results are
available in [11]. The cumulative distribution functions (cdf ) of the failure rate
achieved for 1ooN, 2ooN, 3ooN and majority voting setups are shown in Figure 1.
(a) (b)
(c) (d)
Fig. 1. Cumulative distribution of failure rate for a)1ooN , b)2ooN , c)3ooN , and d) rooN setups.
0.0%
10.0%
20.0%
30.0%
40.0%
50.0%
60.0%
70.0%
80.0%
90.0%
100.0%
0 0. 000 01 -
0.0001
0.0001 -
0.001
0.001 - 0.01 0.01 -0.1 0.1 -1 P r o p o r t i o n o f s y s t e m s w i t h i n a c o n f i g u r a t i o n
Failure rate bands
Cumulative proportion of systems on each failure rate
band for each configuration 1-v
1oo2
1oo3
1oo4
1oo5
1oo6
1oo7
1oo8
1oo9
1oo10
1oo11
1oo12
1oo13
1oo14
1oo15
1oo16
1oo17
1oo18
1oo19
1oo20
1oo21
1oo22
0.0%
10.0%
20.0%
30.0%
40.0%
50.0%
60.0%
70.0%
80.0%
90.0%
100.0%
0 0. 00 00 1 -
0.0001
0.0001 -
0.001
0.001 - 0.01 0.01 -0.1 0.1 -1 P r o p o r t i o n o f s y s t e m s w i t h i n a c o n f i g u r a t i o n
Failure rate bands
Cumulative proportion of systems on each failu re rate
band for each configuration 1-v
2oo2
2oo3
2oo4
2oo5
2oo6
2oo7
2oo8
2oo9
2oo10
2oo11
2oo12
2oo13
2oo14
2oo15
2oo16
2oo17
2oo18
2oo19
2oo20
2oo21
2oo22
0.0%
10.0%
20.0%
30.0%
40.0%
50.0%
60.0%
70.0%
80.0%
90.0%
100.0%
0 0.00001 -
0.0001
0.0001 - 0.001 0.001 - 0.01 0.01 -0.1 0.1 -1
P r o p o r t i o n o f s y s t e m s w i t h i n a c o n f i g u r a t i o n
Failure rate bands
Cumulative proportion of systems on each failure rate band for
each configuration1-v
3oo3
3oo4
3oo5
3oo6
3oo7
3oo8
3oo9
3oo10
3oo11
3oo12
3oo13
3oo14
3oo15
3oo16
3oo17
3oo18
3oo19
3oo20
3oo21
3oo22
3oo23
3oo24
0.0%
10.0%
20.0%
30.0%
40.0%
50.0%
60.0%
70.0%
80.0%
90.0%
100.0%
0 0 .0 00 01 -
0.0001
0.0001 -
0.001
0.001 - 0.01 0.01 -0.1 0.1 -1
P r o p o r t i o n o f s y s t e m s w i t h i n a c o n f i g u r a t i o n
Failure rate bands
Cumulative proportion of systems on each failure rate
band for each configuration1-v
2oo3
3oo5
roo7
roo9
roo11
roo13
roo15
roo17
roo19
roo21
roo23
roo25
roo27
roo29
roo31
roo33
roo35
8/12/2019 Does Malware Detection Improve With Diverse
http://slidepdf.com/reader/full/does-malware-detection-improve-with-diverse 10/13
For part (a), (b) and (c) of Figure 1 the legend shows configurations up to N = 22
or 24, but in fact all 1ooN diverse configurations are contained in the graph. For
1ooN, 2ooN and 3ooN configurations the figures visualize the trend of improving
failure rates as we add more AV products in diverse configurations. The shape of the
distributions for the majority voting setup (part (d) of Figure 1) is different and we see
a decline in the proportion of perfect systems after N=29.
Figure 2 presents the difference in the proportion of the non-perfect detection com-
binations in these three different setups. We can see that to get 90% of all 1ooN sys-
tems to have a perfect detection of all malware with our dataset we need N to be be-
tween 14 and 15 (we can observe this in Figure 2 by following where the x-axis value
14 and 15 meets the y-axis value 1.E-01 for the 1ooN line). To get 90% of 2ooN sys-
tems to detect all malware we need N~20 and 21, whereas N for 3ooN is between 23and 24. This is another way in which an administrator could measure the added cost
of seeking confirmation from 2 or 3 AVs before raising alarms.
Fig. 2. Proportion of diverse systems of size N that have a non-perfect detection rate .
5 Discussion and Comparison of Results With Other Studies
We now discuss and compare the results of the study presented in this paper with
those from our previous study [6-8]. In summary:
In [6-8] despite the generally high detection rates of the AVs, none of themachieved 100% detection rate.
Our observation: we also observe this with the new dataset. Eight of the AVs in
our study had failure rates smaller than 1.E-04 but none of them detected all the in-
stances of malware (on all days) in our study either.
In [6-8] the detection failures were both due to an incomplete signature databases
at the time in which the samples were first submitted for inspection, but also due to
1.E-05
1.E-04
1.E-03
1.E-02
1.E-01
1.E+00
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33
P r o p o
r t i o n o f d i v e r s e s y s t e m s o f s i z e N
t h a t h
a v e a n o n - p e r f e c t d e t e c t i o n r a t e
Number of Diverse AVs (N)
1ooN
proportion
of non-
perfect
systems
2ooN
proportion
of non-
perfect
systems
3ooN
proportion
of non-perfect
systems
8/12/2019 Does Malware Detection Improve With Diverse
http://slidepdf.com/reader/full/does-malware-detection-improve-with-diverse 11/13
regressions in the ability to repeatedly detect malware as a consequence, possibly,
of the deletion of some signatures.
Our observation: we also observe this with the new dataset. Regressions in the de-
tection capabilities were observed even with AVs who are ranked in the top 10 in
our dataset in terms of their detection capability.
In [6-8] considerable improvements in detection rates were observed from employ-
ing diverse AVs: almost 25% of all the diverse pairs, and over 50% of all triplets in
respective 1-out-of-N configurations successfully detected all the malware.
Our observation: we observe improvements in 1-out-of-N configurations though
not as good as those in [6-8]: under 12% of all diverse pairs, and under 24% of all
diverse triplets detected all the malware in 1-out-of-N configurations. To get over
50% of all combinations to detect all the malware we need 6 AVs rather than 3.
In [6-8], no malware caused more than 14 AVs to fail on any given date. Hence perfect detection rates, with their dataset is achieved by using 15 AVs in a 1-out-
of-N configuration.
Our observation: We had malware that failed to be detected by up to 31 AVs in the
new dataset. So to get perfect detection rates with our dataset for all possible in-
stances of a diverse configuration, we need 32 AVs.
In [7] the detection rates were lower for “trigger level detection” (2-out-of-N and
3-out-of-N) and majority voting (r-out-of-N) setups compared with 1-out-of-N but
on average are better than using a single AV.
Our observation: We confirm that these findings are consistent with what we found
previously. Additionally we found that the proportion of perfect majority voting
systems is considerably higher with the new dataset compared with those reported
in [7]: in the new analysis more than 34% of possible combinations of 14oo27 and15oo29 majority voting configurations achieved perfect detection. In [7] the pro-
portion of perfect detection majority voting systems never reached more than 6%
for any of the combinations in that study. This may have implications on the choice
of diverse setups that administrators may use especially if false positives become
an issue. If the detection capability against genuine malware is not (significantly)
hampered from a majority decision by diverse AVs, then majority voting setups
could become more attractive to administrators than 1ooN configurations.
In [6] significant potential gains were observed in reducing the “at risk time” of a
system from employing diverse AVs: even in cases where AVs failed to detect a
malware, there is diversity in the time it takes different vendors to define a signa-
ture to detect a malware.
Our observation: This is also supported by the results in our study (though due to
space limitations we could not elaborate on this here; see [11] for details).
In [6] an empirically derived exponential power law model proved to be a good fit
to the proportion of systems in each simple detection (1-out-of-N) and trigger level
detection (2-out-of-N and 3-out-of-N) diverse setup that had a zero failure rate.
Our observation: we have not explored the modeling aspects in this paper. There
do appear to be some differences between the shapes of the empirical distributions
of the proportion of diverse systems of size N that have a non-perfect detection rate
(see Figure 2) from what is presented in [6]. The empirical distributions in Figure 2
8/12/2019 Does Malware Detection Improve With Diverse
http://slidepdf.com/reader/full/does-malware-detection-improve-with-diverse 12/13
do look like they follow an exponential power law model (possibly of a different
form from that observed in [6]), so further work is needed to check whether the
model outlined in [6] can be generalized for this dataset. A generalized model
would allow a cost-effective prediction of the probability of perfect detection for
systems that use a large number of AVs based on measurements made with sys-
tems that are composed of fewer (say 2 or 3) AVs.
6 Discussion and Conclusions
We reported analysis of results from an empirical study on the possible benefits of
diversity with currently spreading malware and compared and contrasted the findings
with those we reported previously using a different dataset [6-8]. We verified theextent to which the findings previously reported in [6-8] are relevant with more recent
malware. The new results were in general consistent with those reported in [6-8]
though there were differences. The consistent results were:
None of the single AVs achieved perfect detection rates
A number of the AVs regressed in their detection behavior (i.e. failed to detect
malware which they had successfully detected in the past)
Considerable improvements in detection capability when using diversity
As one would expect there is an ordering in the effectiveness of detection capabil-
ity with 1-out-of-N, 2-out-of-N, 3-out-of-N and majority voting diverse setups,
with 1-out-of-N having the best detection rates
Using diverse AVs helps with reducing the “at risk time” of a system
The main differences in the results were:
Despite significant improvements from diversity with 1-out-of-N, 2-out-of-N and
3-out-of-N, they are lower than those reported in [6-8]. For example with 1ooN we
observed that 12% of all diverse pairs and 24% of all diverse triplets detected all
the malware (compared with approximately 25% and 50% respectively in [6-8]).
On the other hand, we observe a much higher proportion of perfect detection ma-
jority voting systems in all setups compared with the results observed in [6-8]
The shape of the empirical distribution of non-perfect detection systems as more
diverse AVs are added seems different from that observed in [6] and should be in-
vestigated further.
The aim of this work is to provide more evidence on the possible benefits of usingdiverse AVs to help with malware detection and therefore help security decision mak-
ers, and other researchers in the field to make more informed, empirically-supported
decisions on the design, assessment and deployment of security protection systems.
The limitations to this work which prevent us from making more general conclu-
sions and hence require further work are as follows. First, we have no data on false
positives. Hence studying the detection capabilities with datasets that allow measure-
ments of false positives in addition to false negative rates will allow us a better analy-
8/12/2019 Does Malware Detection Improve With Diverse
http://slidepdf.com/reader/full/does-malware-detection-improve-with-diverse 13/13
sis of the tradeoffs between the various 1-out-of-N, trigger-level and majority voting
detection setups. Second, we have only tested the detection capability when subject-
ing the AVs to Windows portable executable files. Further studies are needed to
check the detection capability for other types of files e.g. document files, media files
etc. Third, the AV products contain more components than just the signature-based
detection engine which is what the VirusTotal service provides. Further studies are
needed to include the detection capabilities of these products in full. Fourth, due to
lack of space in this paper, we have not explored the modeling aspects and modeling
for prediction. Our next step is to do further work in checking whether a form of the
power law distribution observed in [6] also applies with this dataset.
Current work in progress includes: the patterns with which AVs label the malware
when they detect them. We have started doing this analysis to study the patterns oflabel changes, whether they differ for different products, and whether we can show a
causal link between frequent malware label changes and AVs regressing on their de-
tection capability. In addition, each of the malware in our sample has a different MD5
hashvalue, but it is not clear how similar/different they are. So we are investigating
whether polymorphic malware is identified with the same labels by the AVs and
whether this can aid with diagnosis and recovery from malware infections.
References
1. van der Meulen, M.J.P., S. Riddle, L. Strigini and N. Jefferson. "Protective Wrapping of Off-the-Shelf
Components". in the 4th Int. Conf. on COTS-Based Software Systems (ICCBSS). Bilbao, Spain, p.
168-177, 2005.
2. Strigini, L., "Fault Tolerance against Design Faults", in Dependable Computing Systems: Paradigms,Performance Issues, and Applications, H. Diab and A. Zomaya, Editors, J. Wiley & Sons. p. 213-241,
2005.
3. Oberheide, J., E. Cooke and F. Jahanian. "Cloudav: N-Version Antivirus in the Network Cloud". in
the 17th USENIX Security Symposium, p. 91 – 106, 2008.
4. GFi. Gfimaildefence Suite, last checked 2012, http://www.gfi.com/maildefense/.
5. VirusTotal. Virustotal - a Service for Analysing Suspicious Files, last checked 2013,
http://www.virustotal.com/sobre.html.
6. Bishop, P., R. Bloomfield, I. Gashi and V. Stankovic. "Diversity for Security: A Study with Off-the-
Shelf Antivirus Engines". in the 22nd IEEE International Symposium on Software Reliability Engi-
neering (ISSRE), p. 11-19, 2011.
7. Bishop, P.G., R.E. Bloomfield, I. Gashi and V. Stankovic, "Diverse Protection Systems for Improving
Security: A Study with Antivirus Engines". 2012, City University London: London, UK.
8. Gashi, I., C. Leita, O. Thonnard and V. Stankovic. "An Experimental Study of Diversity with Off-the-
Shelf Antivirus Engines". in the 8th IEEE Int. Symp. on Network Computing and Applications (NCA
2009), p. 4-11, 2009.
9. Schneider, F., "Blueprint for a Science of Cybersecurity". The Next Wave, 19(2): p. 47-57. 2012.
10. Sukwong, O., H.S. Kim and J.C. Hoe, "Commercial Antivirus Software Effectiveness: An Empirical
Study". IEEE Computer, 44(3): p. 63-70. 2011.
11. Cukier, M., I. Gashi, B. Sobesto and V. Stankovic, "Technical report: Does Malware Detection Im-
prove with Diverse Antivirus Products? An Empirical Study", 2013,
http://www.csr.city.ac.uk/people/ilir.gashi/SAFECOMP2013/